JL1-CC&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering

Ziyuan Liu, Ruifei Zhu, Ouqiao Ma, Yuantao Gu · 2026

Read JL1-CC&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering by Ziyuan Liu, Ruifei Zhu, Ouqiao Ma, Yuantao Gu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-31745.txt.

Introduction

Liu, Zhu, Ma and Gu extend a satellite change-detection dataset so models must say what changed, not only where. Standard change detection outputs a binary mask over two images of the same place. It stays silent about the kind of change and the reasons behind it. The team builds on 5,000 image pairs from the Jilin-1 satellite.

Two annotation layers are added. One holds 17,021 checked captions describing land-cover transformations in words. The other holds 20,060 question-answer pairs spread over eight question types. Annotations came from a three-stage pipeline: model generation, vision-grounded judging, then human expert verification. Together with the original masks, the layers form one benchmark for multi-task change understanding.

Catalog record

CREATORZiyuan Liu, Ruifei Zhu, Ouqiao Ma, Yuantao Gu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.31745

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

JL1-CC&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering Ziyuan Liu, Ruifei Zhu, Ouqiao Ma, Yuantao Gu Abstract Remote sensing change detection (CD) traditionally focuses on pixel-level binary segmentation, which identifies where changes occur but neither what nor why . To bridge this semantic gap, we introduce JL1-CC&QA, a multi-task benchmark that extends the JL1-CD dataset with two complementary annotation layers: change captioning (CC) and change question answering (QA). Built upon 5,000 bi-temporal image pairs acquired by the Jilin-1 satellite at 0.5–0.75 m ground sample distance, the benchmark comprises: (i) JL1-CC , providing 17,021 quality-verified captions that describe diverse land-cover transformations; and (ii) JL1-QA , offering 20,060 question–answer pairs across eight question types, enabling fine-grained, interactive interrogation of surface changes. All annotations are produced via a three-stage pipeline consisting of multi-modal large language model (LLM) generation, vision-grounded LLM judging, and human expert verification. We hope that JL1-CC&QA, as a benchmark unifying binary change masks, change captions, and change-oriented QA over the same image set, will serve as a valuable resource for the community to advance multi-task change understanding in remote sensing. The dataset is available at https://github.com/circleLZY/JL1-CD . I Introduction Remote sensing change detection (CD) aims to identify surface changes from multi-temporal imagery acquired over the same geographic area, serving as an important technology for urban planning, environmental monitoring, disaster response, and resource management [ 79 , 2 ] . Over the past decade, the rapid growth of both Earth observation data and deep learning methods has propelled CD into one of the most active research frontiers in remote sensing. The predominant formulation of CD is binary change detection (BCD), which classifies each pixel as either changed or unchanged . A large number of BCD benchmarks have been established, spanning diverse geographic contexts, spatial resolutions, and change categories (see Table I for a comprehensive summary) [ 5 , 8 , 22 , 51 , 47 , 108 , 78 , 13 , 121 , 85 , 90 , 89 , 93 , 69 , 53 , 40 , 77 ] . While the vast majority of these benchmarks adopt the bi-temporal setting and have driven extensive algorithmic development [ 24 , 13 , 121 , 30 , 12 , 4 , 16 ] , the community has also explored single-temporal CD that exploits pretrained semantic representations to infer changes from unpaired imagery [ 131 , 17 , 137 ] , as well as multi-temporal monitoring that tracks continuous change trajectories from dense satellite time series [ 103 , 105 , 94 , 129 ] . Despite these differences, all of the above approaches produce pixel-level masks that encode where change occurred but remain silent on what changed or why . TABLE I: Comprehensive Summary of Representative Remote Sensing Change Detection Datasets Dataset Year Task Modal. Source GSD Size (px) Scale Temp. Change Elements SZTAKI [ 5 ] 2009 BCD Opt. Aerial 1.5 m 952\!\times\!640 13 Bi Buildings, forest, foundations AICD [ 8 ] 2011 BCD Opt. Synthetic — 800\!\times\!600 1000 Bi Synthetic (buildings, vehicles) ABCD [ 34 ] 2017 BDA Opt. Aerial 0.4 m 128\!\times\!128 8506 Bi Building damage MtS-WH [ 110 ] 2017 BCD Opt. IKONOS 1 m 7200\!\times\!6000 1 Bi Parking, water, bldg., farm, industrial OSCD [ 22 ] 2018 BCD MS Sentinel-2 10 m 600\!\times\!600 24 Bi Buildings, roads CDD [ 51 ] 2018 BCD Opt. GE VHR 256\!\times\!256 16000 Bi Seasonal, vehicles, bldg., trees WHU-CD [ 47 ] 2018 BCD Opt. Aerial 0.2 m 15354\!\times\!32507 1 Bi Buildings HRSCD [ 23 ] 2019 SCD Opt. Aerial (IGN) 0.5 m 10000\!\times\!10000 291 Bi Artificial, agriculture, forest, wetland, water River [ 108 ] 2019 BCD HS Airborne — 463\!\times\!241 1 Bi Water Hyper. CDD [ 78 ] 2019 BCD HS Airborne — Varies 3 Bi Crop transition xBD [ 35 ] 2019 BDA Opt. Satellite VHR 1024\!\times\!1024 22068 Bi Building damage LEVIR-CD [ 13 ] 2020 BCD Opt. GE 0.5 m 1024\!\times\!1024 637 Bi Buildings LEVIR-CD+ [ 13 ] 2020 BCD Opt. GE 0.5 m 1024\!\times\!1024 985 Bi Buildings DSIFN [ 121 ] 2020 BCD Opt. GE VHR 512\!\times\!512 3940 Bi Roads, bldg., farmland, water CD_Data_GZ [ 85 ] 2020 BCD Opt. GE VHR 256\!\times\!256 4000+ Bi Water, road, farmland, bare soil, forest, bldg., ships Optical-SAR-CD [ 88 ] 2021 BCD Cross Opt.+SAR 10 m 600\!\times\!600 3 Bi Mixed land cover SYSU-CD [ 93 ] 2021 BCD Opt. Aerial 0.5 m 256\!\times\!256 20000 Bi Bldg., road, veg., water, marine construction S2Looking [ 90 ] 2021 BCD Opt. Satellite 0.5–0.8 m 1024\!\times\!1024 5000 Bi Buildings HTCD [ 89 ] 2021 BCD Opt. Sat.+UAV VHR Varies 1293 Bi Buildings, roads, urban features SECOND [ 115 ] 2021 SCD Opt. Aerial 0.5–3 m 512\!\times\!512 4662 Bi Ground, tree, low veg., water, bldg., playground ChangeOS [ 134 ] 2021 BDA Opt. Satellite VHR Varies 2568 Bi No damage / minor / major / destroyed ACDA [ 43 ] 2021 BCD HS Air./Sat. — Varies 4 Bi Anomalous spectral changes SpaceNet 7 [ 105 ] 2021 BCD Opt. Satellite 4 m 1024\!\times\!1024 101 AOIs Multi Buildings WHU-OPT-SAR [ 60 ] 2022 BCD Cross GF-2+S1 1–10 m Varies 2695 Bi Farmland, urban, rural, water, forest, road CLCD [ 69 ] 2022 BCD Opt. GF-2 0.5–2 m 512\!\times\!512 600 Bi Cropland Hi-UCD [ 102 ] 2022 SCD Opt. Aerial 0.1 m 1024\!\times\!1024 40800+ Multi Water, grass, forest, bare soil, bldg., greenhouse, road, bridge Landsat-SCD [ 118 ] 2022 SCD MS Landsat 30 m 256\!\times\!256 8468 Bi Desert, farmland, bldg., water MSBC [ 53 ] 2022 BCD Multi GF-2+S1+S2 0.8–10 m Varies 4838 Bi Buildings DynamicEarthNet [ 103 ] 2022 SCD MS PlanetScope 3 m 1024\!\times\!1024 75 AOIs Multi Impervious, agriculture, forest, wetland, soil, water, snow BANDON [ 82 ] 2023 SCD Opt. Aerial VHR 2048\!\times\!2048 2283 Bi Buildings 3DCD [ 80 ] 2023 BCD Opt. Aerial VHR 512\!\times\!512 472 Bi Buildings, roads, bridges CAU-Flood [ 39 ] 2023 BCD Cross Opt.+SAR 10 m 256\!\times\!256 2016 Bi Flood inundation WUSU [ 94 ] 2023 SCD Opt. GF-2 0.8 m 512\!\times\!512 10000+ Multi Bldg., road, cropland, forest, grass, river, lake, excavation, bare soil RescueNet [ 87 ] 2023 BDA Opt. UAV VHR 3000\!\times\!4000 2396 Single Bldg., water, road, vehicle, tree, pool EGY-BCD [ 40 ] 2023 BCD Opt. Satellite 0.25 m 256\!\times\!256 6091 Bi Buildings CropSCD [ 68 ] 2024 SCD Opt. GF-2 VHR 256\!\times\!256 4202 Bi Water, forest, plantation, grassland, impervious, greenhouse, road, bare soil ClearSCD [ 101 ] 2024 SCD Opt. Various 0.5–3 m 512\!\times\!512 4662 Bi Ground, tree, low veg., water, bldg., playground QQB [ 97 ] 2024 BDA Cross SAR+Opt. VHR 256\!\times\!256 3200 Bi Buildings TSCD [ 129 ] 2024 BCD Opt. WorldView-2 0.5 m 256\!\times\!256 5140 Multi Buildings, greenhouses JL1-CD [ 77 ] 2025 BCD Opt. JL-1 0.5–0.75 m 512\!\times\!512 5000 Bi Bldg., road, PV, hardened, forest, grass, cropland, water PRO-HRSCD [ 31 ] 2025 SCD Opt. Aerial (IGN) 0.5 m 10000\!\times\!10000 291 Bi Artificial, agriculture, forest, wetland, water TripleS [ 98 ] 2025 SCD Opt. Various 0.5–3 m 512\!\times\!512 4662+ Bi Bare soil, water, bldg., farmland, veg., road BRIGHT [ 15 ] 2025 BDA Cross SAR+Opt. 0.3–1 m 1024\!\times\!1024 4538 Bi Buildings CDFNet [ 106 ] 2025 BDA Opt.

Read the complete text